Align training trajectory inputs with original open-domain SFT settings

cot_trajectory_generation/glm5.2/Multi-subject-RLVR/SOURCE.md

@@ -4,7 +4,5 @@
 - Source: https://huggingface.co/datasets/virtuoussy/Multi-subject-RLVR
 - The published dataset includes a system message; it is preserved verbatim in
   `official_dataset_system_prompt.txt` for provenance.
-- Applied input: the exact Qwen training system prompt in
-  `training_system_prompt.txt`. The official dataset prompt is not used for
-  generation because the actual latent-SFT training loader replaces it.
-
+- Applied input: the published system message followed by the original user
+  question. No additional boxed-answer prompt is added.

cot_trajectory_generation/glm5.2/Multi-subject-RLVR/pending.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:75619490a9cb187d1bf7345b94e183778e418ea7ac8a4f67b2d33e56ef39e47a
-size 23977000
+oid sha256:a679bd214b6d8d1a87a399a718bf0f3e9d131c9a166f589d67af2288fa6a5108
+size 24905646

cot_trajectory_generation/glm5.2/Multi-subject-RLVR/training_system_prompt.txt

@@ -1 +0,0 @@
-Please reason step by step, and put your final answer within \boxed{}.

cot_trajectory_generation/glm5.2/README.md

@@ -16,13 +16,18 @@ This directory contains only the seven datasets currently in scope.
 
 ## Input alignment
 
-- Training datasets reproduce the Qwen branch of
-  `src/stage1/data.py` at commit
-  `a50b4794c6743b936c41b317b226148919a9afd7`.
+- Training datasets reproduce their original open-domain SFT query setting.
+- Multi-subject-RLVR retains its published system message.
+- ARC, TriviaQA, and WebInstruct-verified have no source system message in the
+  local source records, so their queries contain only the complete user input.
+- No training dataset is given an additional `\boxed{}` instruction.
 - MMLU-Pro reproduces the official category-specific few-shot user prompt.
 - SuperGPQA reproduces the official reasoning-model zero-shot user prompt.
 - Neither validation benchmark is given an invented system role.
 
+The current Qwen branch of `src/stage1/data.py` still hard-codes a boxed system
+prompt. It must be updated before training so that the runtime input matches
+these dataset-native queries.
+
 Each dataset directory contains `pending.jsonl.gz`. Existing validation
 trajectories are retained separately in `completed.jsonl.gz`.
-

cot_trajectory_generation/glm5.2/WebInstruct-verified/SOURCE.md

@@ -3,6 +3,4 @@
 - Role: training
 - Source: https://huggingface.co/datasets/TIGER-Lab/WebInstruct-verified
 - Official dataset system prompt: none published
-- Applied input: the exact Qwen training system prompt in
-  `training_system_prompt.txt`, followed by the original question.
-
+- Applied input: one user message containing the original question.

cot_trajectory_generation/glm5.2/WebInstruct-verified/pending.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:89592133cf135d927f48c7344c169f7c0378ef389b6b0b133736644754e25c7e
-size 23687858
+oid sha256:9d39ecf9d0ab12a3c41dcfa18066d65c6ba41b531d655dc38d0a19b3b2eb632f
+size 23363726

cot_trajectory_generation/glm5.2/WebInstruct-verified/training_system_prompt.txt

@@ -1 +0,0 @@
-Please reason step by step, and put your final answer within \boxed{}.

cot_trajectory_generation/glm5.2/ai2-arc-ARC-Challenge/SOURCE.md

@@ -4,6 +4,4 @@
 - Source: https://huggingface.co/datasets/allenai/ai2_arc
 - Configuration: `ARC-Challenge`
 - Official dataset system prompt: none published
-- Applied input: the exact Qwen training system prompt in
-  `training_system_prompt.txt`, followed by the complete question and options.
-
+- Applied input: one user message containing the complete question and options.

cot_trajectory_generation/glm5.2/ai2-arc-ARC-Challenge/pending.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8957ee6765775c7fa74fe4453a8bc95ee55c52f3aa9c9fbc9e15d84c43e0b2fd
-size 126064
+oid sha256:241b48a394ddbc1c6a7cf3eb146758cab92a84589a524c858cd5c6bbf60151bd
+size 123866

cot_trajectory_generation/glm5.2/ai2-arc-ARC-Challenge/training_system_prompt.txt

@@ -1 +0,0 @@
-Please reason step by step, and put your final answer within \boxed{}.

cot_trajectory_generation/glm5.2/ai2-arc-ARC-Easy/SOURCE.md

@@ -4,6 +4,4 @@
 - Source: https://huggingface.co/datasets/allenai/ai2_arc
 - Configuration: `ARC-Easy`
 - Official dataset system prompt: none published
-- Applied input: the exact Qwen training system prompt in
-  `training_system_prompt.txt`, followed by the complete question and options.
-
+- Applied input: one user message containing the complete question and options.

cot_trajectory_generation/glm5.2/ai2-arc-ARC-Easy/pending.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7a040413b595f639b4986f15d2b86259a876a0323781c9faa6b55ef93d9f4c99
-size 240101
+oid sha256:3e81a5184b3959742df4c8b8d3566574888dcc7f2c0f2a1a988ad278c7b1e459
+size 235469

cot_trajectory_generation/glm5.2/ai2-arc-ARC-Easy/training_system_prompt.txt

@@ -1 +0,0 @@
-Please reason step by step, and put your final answer within \boxed{}.

cot_trajectory_generation/glm5.2/manifest.json

@@ -7,12 +7,6 @@
     "postprocess_think_tokens": true,
     "postprocess_box_if_required": true
   },
-  "training_input_source": {
-    "code_commit": "a50b4794c6743b936c41b317b226148919a9afd7",
-    "code_path": "src/stage1/data.py",
-    "model_family": "qwen",
-    "system_prompt": "Please reason step by step, and put your final answer within \\boxed{}."
-  },
   "counts": {
     "ai2-arc-ARC-Easy": {
       "total": 2111,
@@ -49,5 +43,19 @@
       "pending": 26385,
       "completed": 144
     }
+  },
+  "training_input_policy": {
+    "policy": "follow each open-domain SFT dataset's original query setting",
+    "datasets_with_system_prompt": {
+      "Multi-subject-RLVR": "You are a chatbot who can solve problems. Please solve the following problem and give your thought process. Before giving the final result, you should output \"Therefore, the answer is\", and then give your final answer."
+    },
+    "datasets_without_system_prompt": [
+      "ai2-arc-ARC-Easy",
+      "ai2-arc-ARC-Challenge",
+      "trivia-qa",
+      "WebInstruct-verified"
+    ],
+    "boxed_prompt_added": false,
+    "training_code_note": "The current Qwen branch in src/stage1/data.py still hard-codes a boxed system prompt and must be changed before aligned training."
   }
 }

cot_trajectory_generation/glm5.2/trivia-qa/SOURCE.md

@@ -4,6 +4,4 @@
 - Source: https://github.com/mandarjoshi90/triviaqa
 - Configuration used locally: `rc.nocontext`
 - Official dataset system prompt: none published
-- Applied input: the exact Qwen training system prompt in
-  `training_system_prompt.txt`, followed by the original question.
-
+- Applied input: one user message containing the original question.

cot_trajectory_generation/glm5.2/trivia-qa/pending.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e18533eb141f72ffe31cd2a4daca716ed1bd6d0b33a354886c760272d86c7cd5
-size 7339982
+oid sha256:6bee048e9bf7f3e7adc13888464540ce545c669bfcfe5371b36aa338f2000b6f
+size 7124280

cot_trajectory_generation/glm5.2/trivia-qa/training_system_prompt.txt

@@ -1 +0,0 @@
-Please reason step by step, and put your final answer within \boxed{}.